The six audits that have to pass before a rating ships
A rating system you can edit by hand will be edited by hand, usually at 11pm, usually to make a build go green. The rule here is the opposite. The catalog is recomputed from signals. Six checks have to pass. If they do not, nothing ships. The instruction in the pipeline docs is not to weaken the check. It is to fix the data.
These are not style lints. Each one exists because a specific lie got into a shipped build.
1. Within-album inversions
ratings:audit. If track A is rated above track B on the same album, A needs a reason: more streams, a better chart peak, a certification, views, or owner-approved cap relief. A deep cut outranking a hit with nothing behind it fails. Justified inversions are counted and reported. They are not silently allowed to look like a green run.
This is the audit that would have caught the Chili Peppers collision earlier if I had been looking at albums instead of rows. I wrote that up as the bucket bug. The check now runs on every build.
2. Signal integrity
ratings:audit:signals. Duplicate Spotify track IDs anywhere in the catalog are a hard fail. Reprise and interlude hijacks are a hard fail. Weak albums get reported so I cannot pretend a five-track EP with one matched song is a complete era. This is the gate in front of Hey Ya! (Reprise).
3. Stream order and spread cliffs
ratings:audit:streams. Within an album, stream order and rating order have to reconcile through the same inversion rule. Absolute-ladder drift is flagged. Spreads that explode past the compressed 8–16 point target get reported so a twenty-track record cannot look half unlistenable because of interludes.
It also bans a specific cheat: no unpinned track whose true total streams sit between about 350 and 450 million is allowed to rate 90 or above. That is the Wolves / Big Energy band. Era caps are earned; they are not a magnet for every song of that year.
4. Named regressions
ratings:audit:regressions. A short list of cases that already burned me, checked by name. Hey Ya! (Reprise) must stay far below Hey Ya! and out of the 90s. Mariah Carey's Without You must not float into the mid-90s. Drake's Started From the Bottom must stay at 86 or above and close to From Time. Wolves and Big Energy must stay out of 90. Beyoncé's COZY and Ariana Grande's the light is coming must not fall off a cliff into the 60s despite ~166 million streams.
A regression suite is how you stop fixing one album by breaking another. It is also how you keep a conversation with yourself honest. If I want Wolves at 90, I have to delete the check in public, not comment it out locally.
5. Bake drift
ratings:audit:drift. Every rating stored in the artist seed files is recomputed from the current signals. If any baked number disagrees with a fresh run, the build fails. You cannot hand-edit a 94 into a 96 and leave it to rot. Pins exist, they are a separate JSON file, there are a few dozen, and they are owner-approved. A pin is an admission. Drift is how I keep the admissions from spreading.
6. Signal truth
ratings:audit:truth. The algorithm's idea of "effective streams" has to match the match-ledger. Reasons files have to exist. This is the audit that makes a green run mean the inputs are the inputs, not a leftover cache from a matcher I already replaced.
The 11pm rule
All six run, in that order, as ratings:audit:all, and they are a prerequisite of the production build. The temptation is always the same: one inversion on an album I do not care about, one reprise that is "close enough," one pin nobody will notice. Each of those is how the Chili Peppers landed on 94/94/94. The docs say do not weaken the check to go green. Fix the data. I put it there because I am the person it is aimed at.
Player-facing version of what survives this gauntlet: the numbers you see in Classic, and the numbers you do not see in Expert until the end. How those numbers become a chart debut is the score note.
Play #1 Debut